LINEARIZING E-CLASS POWER AMPLIFIER BY USING
MEMORYLESS PRE-DISTORTION
by
Tunir Dey
A Thesis
Submitted to the Faculty of Purdue University
In Partial Fulfillment of the Requirements for the degree of
Master of Science in Engineering
School of Electrical & Computer Engineering
Fort Wayne, Indiana
December 2018
2
THE PURDUE UNIVERSITY GRADUATE SCHOOL
STATEMENT OF COMMITTEE APPROVAL
Dr. Abdullah Eroglu, Chair
Department of Electrical and Computer Engineering
Dr. Carlos Pomalaza-Raez
Department of Electrical and Computer Engineering
Dr. Carolyn Stumph
Doermer School of Business
Approved by:
Dr. Guoping Wang
Head of the Graduate Program
3
ACKNOWLEDGMENTS
First, I would like to thank my thesis advisor Dr. Abdullah Eroglu Emeritus Faculty at Purdue
University, Fort Wayne. Prof. Dr. Eroglu’s office door was always open whenever I had a question
about my research or facing a trouble. He had continually showed me the right direction to
complete this work whenever I needed it.
I also would like to thank an expert who walked me through a part of this research project:
[National Instrument LabVIEW Engineer, Sundaram Raghuraman for his help to make me
understand DPD analysis]. Without their devoted participation and input, part of the verification
work could not have been reach so far.
Finally, I must appreciate my parents and thankful for providing me support and continuous
encouragement throughout my years of study. This achievement would not have been feasible
without them.
Thank you.
Author
Tunir Dey
4
TABLE OF CONTENTS
LIST OF TABLES .......................................................................................................................... 6
LIST OF FIGURES ........................................................................................................................ 7
LIST OF ABBREVIATIONS ......................................................................................................... 9
ABSTRACT .................................................................................................................................. 10
INTRODUCTION .............................................................................................. 12
1.1 Wireless Link & Power Amplifiers .................................................................................. 12
1.2 Importance of Linearity & Non-Constant Envelope ......................................................... 13
1.3 Non- Linearity Effect of Power Amplifiers ...................................................................... 13
1.4 AM/AM & AM/PM Distortions ....................................................................................... 16
1.5 Thesis Outline ................................................................................................................... 17
INTRODUCTION TO POWER AMPLIFIER AND ITS EFFICIENCY .......... 18
2.1 Power Amplifier Classes and Efficiency .......................................................................... 18
POWER AMPLIFIER DESIGN AND IT’S NON-LINEARITY ....................... 22
3.1 13.56MHz E- Class Power Amplifier Design and Simulation ......................................... 22
3.1.1 Analytical Analysis .................................................................................................... 22
3.1.2 Achievable Waveforms ............................................................................................. 23
3.2 Effects of Power Amplifier Nonlinearity .......................................................................... 26
3.2.1 Harmonic Generation ................................................................................................. 26
3.2.2 Intermodulation Distortion ........................................................................................ 26
3.2.3 Spectral Regrowth ..................................................................................................... 26
3.2.4 Cross Modulation....................................................................................................... 27
3.2.5 Desensitization ........................................................................................................... 28
LINEARIZATION TECHNIQUES .................................................................... 29
4.1 Linearization Techniques .................................................................................................. 29
4.1.1 Feedback .................................................................................................................... 29
4.1.2 Feedforward ............................................................................................................... 30
4.1.3 Predistortion ............................................................................................................... 30
BASEBAND PREDISTORTION MODELLING OF NON-LINEAR SYSTEMS
…………………………………………………………………………………..33
5
5.1 Introduction ....................................................................................................................... 33
5.2 Bandpass models versus baseband models ....................................................................... 33
5.3 The Quasi-Memoryless Case ............................................................................................ 35
5.3.1 AM/AM- and AM/PM-Conversion ........................................................................... 36
5.3.2 Frequency-Domain Representation ........................................................................... 37
5.4 Complex Baseband Modeling with Volterra Series .......................................................... 37
5.4.1 Time-Domain Representation .................................................................................... 38
5.5 Comparison of the Model Structure Performances of the Predistorters ........................... 41
MEMORY POLYNOMIAL DIGITAL PREDISTORTION .............................. 43
6.1 Theory ............................................................................................................................... 43
6.2 Inverse Structures.............................................................................................................. 44
6.2.1 pth-order Inverse Method .......................................................................................... 44
6.2.2 Indirect Learning Architecture .................................................................................. 45
6.2.3 Direct learning architecture ....................................................................................... 45
6.3 The Improved Pre-Distortion Algorithm for Indirect Learning Architecture (ILA) ........ 46
6.4 Estimation Algorithms ...................................................................................................... 47
6.4.1 Least Mean-Square Algorithm (LMS) ....................................................................... 47
6.4.2 Recursive Least-Squares Algorithm (RLS) ............................................................... 48
SIMULATION RESULTS ................................................................................. 49
7.1 Introduction ....................................................................................................................... 49
7.2 Modeling and Simulating the Power Amplifier (Saleh Model) ........................................ 49
7.3 Deriving DPD Coefficient: ............................................................................................... 50
7.3.1 Memory-Polynomial Model implementation using Matlab Simulink....................... 51
7.4 Evaluating the Static-Coefficient DPD Design................................................................. 54
7.5 Extend static DPD design to adaptive ............................................................................... 57
CONCLUSION ................................................................................................... 62
FUTURE WORK ................................................................................................ 64
REFERENCES.................................................................................................. 65
6
LIST OF TABLES
Table 1.1 Amplifier’s Efficiency, Layout Area and Distortion .................................................... 21
Table 1.2 Equation 7.1 Can Be Implemented Using Either MATLAB Code or Blocks in
Simulink ........................................................................................................................................ 55
7
LIST OF FIGURES
Figure 1.1 Block Diagram of Modern Transmitter [8] ................................................................. 12
Figure 1.2 Sample 256 QAM, OFDM signal [8] .......................................................................... 13
Figure 1.3 Input Power vs Output Power For 1 dB Compression Point [8] ................................. 14
Figure 1.4 Comparison of Two Different Input Signals With Different Power Levels [8] .......... 15
Figure 1.5 AM/PM Conversion .................................................................................................... 16
Figure 1.6 AM/AM and AM/PM Curves [8] ................................................................................ 17
Figure 2.1 Class A Amplifier Design and Output......................................................................... 18
Figure 2.2 Class B Amplifier Design and Output ......................................................................... 19
Figure 2.3 Class AB Amplifier Design and Output ...................................................................... 19
Figure 2.4 Class C Amplifier Design and Output ......................................................................... 20
Figure 2.5 Amplifier’s Efficiency and Conduction Angle ............................................................ 21
Figure 3.1 13.56 MHz E- Class Power Amplifier Schematic Design .......................................... 22
Figure 3.2 Class-E Waveforms With Ideal Components And Under Optimal Switching
Conditions ..................................................................................................................................... 24
Figure 3.3 13.56MHz PA Output .................................................................................................. 25
Figure 3.4 Spectral Regrowth ....................................................................................................... 27
Figure 4.1 Feedback Linearization [4] .......................................................................................... 29
Figure 4.2 Feedforward Linearization [4] ..................................................................................... 30
Figure 4.3 Predistortion Method of Linearization [4] ................................................................... 31
Figure 4.4 a) RF predistortion, b) IF predistortion, c) Baseband predistortion [4]....................... 32
Figure 5.1 Baseband Model of a Quasi-memoryless System, which is Composed of Two Static
Non-linear Functions Described By The AM/AM Conversion |v(a)| and AM/PM Conversion arg
{v(a)}. [9] ...................................................................................................................................... 37
Figure 5.2 Complex Baseband Time-Domain Output Signals of a 13.56 MHz RF PA, a
Memoryless PA Model. ................................................................................................................ 41
Figure 6.1 Block Diagram of Digital Predistortion Scheme [4] ................................................... 43
Figure 6.2 The Indirect Learning Architecture (ILA) For Non-linear Compensation. ................. 45
Figure 6.3 Direct Learning Architecture ....................................................................................... 46
Figure 7.1 The E-class PA Measuring Model ............................................................................... 49
8
Figure 7.2 The Structure of The Saleh Memoryless Non-linearity [6] ......................................... 50
Figure 7.3 The PA Output Shows 26.49 dB of Gain at The Expense of Significant Spectral
Regrowth ....................................................................................................................................... 50
Figure 7.4 DPD Derivation and Implementation [6] .................................................................... 51
Figure 7.5 Basics of Memory Effect [4] ....................................................................................... 52
Figure 7.6 Equation 7.1 Rearranged As a System of Linear Equations in Matrix Form [6] ........ 53
Figure 7.7 The Real and Imaginary Components of The Derived DPD’s Complex Coefficients.54
Figure 7.8 DPD Verification Model ............................................................................................. 56
Figure 7.9 PA Behavior With DPD (1) ......................................................................................... 56
Figure 7.10 PA Behavior With DPD (2) ....................................................................................... 57
Figure 7.11 Our Adaptive DPD Test Bench ................................................................................. 58
Figure 7.12 Coefficient Computation Subsystem Model [6] ........................................................ 58
Figure 7.13 The Nonlinear Prod Subsystem Implements The Non-Linear Multiplies [6] ........... 59
Figure 7.14 The LMS-Based Coefficient Computation is Relatively Simple In Terms of The
Required Resources [6] ................................................................................................................. 59
Figure 7.15 The RPEM-Based Coefficient Computation is Far More Complex Than The LMS
[6] .................................................................................................................................................. 60
Figure 7.16 Adaptive DPD Design Output Behavior ................................................................... 61
9
LIST OF ABBREVIATIONS
DPD Digital Pre-distortion
AM/PM Amplitude Modulation/ Phase Modulation
GaAs Gallium Arsenide
GaN Gallium Nitride
OFDM Orthogonal Frequency Division Multiplexing
W-CDMA Wideband Code Division Multiple Access
PAPR Peak to Average Power Ratio
CEPT European Conference of Postal and Telecommunications Administrations
FET Field-Effect Transistor
IGBT Insulated-Gate Bipolar Transistor
MOSFET Metal-Oxide-Semiconductor Field-Effect Transistor
DSP Digital Signal Processing
PWM Pulse Width Modulation
PA Power Amplifier
RMS Root Mean Square
ACPR Adjacent Channel Power Ratio
IF Intermediate Frequency
EVM Error Vector Magnitude
DUT Device Under Test
ILA Indirect Learning Architecture
DLA Direct Learning Architecture
LMS Least Mean-Square
RLS Recursive Least-Squares
AWGN Additive White Gaussian Noise
RPEM Recursive Predictor Error Method
ILC Iterative Learning Control
10
ABSTRACT
Author: Dey, Tunir Initial. TD
Institution: Purdue University Fort Wayne
Degree Received: December 2018
Title: Linearizing E- Class Power Amplifier by Using Memoryless Pre-Distortion
Committee Chair: Abdullah Eroglu
Radio Frequency Power Amplifiers (PA) are essential components of wireless systems and
nonlinear in a permanent way. So, high efficiency and linearity at a time are imperative for power
amplifiers. However, it is hard to obtain because high efficiency Power Amplifiers are nonlinear
and linear Power Amplifiers have poor efficiency. To meet both linearity and efficiency, the
linearization techniques such as Digital Predistortion (DPD) has arrested the most attention in
industrial and academic sectors due to provide a compromising data between efficiency and
linearity. This thesis proposed on digital predistortion techniques to control nonlinear distortion in
radio frequency transmitters.
By using predistortion technique, both linearity and efficiency can obtain. In this thesis a new
generic Saleh model for use in memoryless nonlinear power amplifier (PA) behavioral modelling
is used. The results are obtained by simulations through MATLAB and experiments. We explore
the baseband 13.56 MHz Power Amplifier input and output relationships and reveal that they
apparent differently when the Power Amplifier shows long-term, short-term or memory less effects.
We derive a SIMULINK based static DPD design depend on a memory polynomial. A polynomial
improves both the non-linearity and memory effects in the Power Amplifier. As PA characteristics
differs from time to time and operating conditions, we developed a model to calculate the
effectiveness of DPD. We extended our static DPD design model into an adaptive DPD test bench
using Indirect Learning Architecture (ILA) to implement adaptive DPD which composed of DPD
subsystem and DPD coefficient calculation. By this technique, the output of PA achieves linear,
amplitude and phase distortions are eliminated, and spectral regrowth is prevented.
The advanced linearity performance executed through the strategies and methods evolved on
this thesis can allow a higher usage of the capability overall performance of existing and
emerging exceptionally performance PAs, and therefore an anticipated to have an effect in future
wireless communication systems.
Keywords
11
PA Efficiency, Behavioral model, Digital Pre-Distortion, Power Amplifier, Nonlinear, Volterra
series, AM/ AM, AM/ PM, Static DPD, Adaptive DPD.
12
INTRODUCTION
1.1 Wireless Link & Power Amplifiers
Figure 1.1 Block Diagram of Modern Transmitter [8]
Power amplifiers (PA) are one of the essential block in wireless systems [8], which plays a major
role to overcome the loss between the transmitter and receiver. Figure 1.1 shows a block diagram
of modern transmitter [8]. To achieve the adequate transmission power, Power Amplifiers are
essentials in wireless or wireline systems. To get high efficient level we can increase the input
power of power amplifier. However, PAs are non-linear when the input power is high, which
builds a trade-off between efficiency and linearity. The last output in wireless system includes
power amplifier to send the signal to the antenna. Depending on the frequency band, the power
output and the desired efficiency, designers can select PA’s fabricated with a wide range of
technologies like: Gallium Arsenide (GaAs), Gallium Phosphide, Silicon Germanium, CMOS or
Gallium nitride (GaN). Moreover, high speed of communication in wireless industry has
controlled by 3GPP like Wideband Code Division Multiple Access (W-CDMA) which is using
unstable envelope techniques. These techniques have high peak to average power ratio (PAPR)
when increase the frequency spectrum efficiency. To overcome this non-linearity - digital
predistortion (DPD) is one of the widely accepted techniques. In this paper we have examined and
implemented different DPD techniques.
13
1.2 Importance of Linearity & Non-Constant Envelope
Now-a-days in wireless industry communication techniques used in 3G (W-CDMA) and 4G
(OFDM) both uses unstable envelope techniques [8]. In figure 1.2 a 256 QAM OFDM signal
sample has shown. As we said these techniques also has high peak to average power ratio which
makes a conflict between power efficiency and linearity. In this paper we discussed about the
conflict and the PA efficiency. And, importance of the linearity will be discussing in this section.
When a Power Amplifier is behaving non-linearly, it starts to create harmonic distortion in the
output which comes with a frequency spectral growth. However, due to strict frequency
allocations this needs to be avoided.
Figure 1.2 Sample 256 QAM, OFDM signal [8]
1.3 Non- Linearity Effect of Power Amplifiers
So far, we talked about the significance of linearity, but PA non- linearity has not been discussed
yet. In this section the reasons of PA’s non-linearity will be discussed. Due to high frequency PA
in communication systems, nonlinear behavior happens which degrades signal. This is unlikely to
the signals that pass through it and appears Harmonic Distortion, Gain compression, Inter-
modulation distortion etc. A Power Amplifier should be driven in high power mode to obtain a
high-power efficiency but after a definite amount of input power level each Power Amplifier tends
14
to hit the compression region. A power amplifier has promising gain while drive into low power
level but for high power levels PA goes in to saturation and the gain does not meet as expected
like low power level which means gain decreased, the point where the gain decreased 1dB [8].
From its constant value is called 1dB compression point.
Figure 1.3 Input Power vs Output Power For 1 dB Compression Point [8]
This cutting in time domain has a spectral growth effect in frequency domain. Figure 1.4 (a) and
figure 1.4 (b) are showing the effect of cutting, in figure 1.4(a) is shown a sin wave of 1MHz
frequency and its frequency domain response. In figure 1.4(b) same wave but with five percent
cutting and its frequency response has shown as well. When there is cutting, the signal grows in
frequency domain. Additionally, in figure 1.4(c) [8] two different input signals comparison has
shown. Here, Intermodulation products are more essentials and by using filters rest of the
harmonics can be restrain.
15
(a)
(b)
(c)
Figure 1.4 Comparison of Two Different Input Signals With Different Power Levels [8]
16
1.4 AM/AM & AM/PM Distortions
PA non-linearity can be obtained by amplitude to amplitude (AM/AM) conversion and amplitude
to phase (AM/PM) conversion. AM/AM represents a digression from a constant gain when input
power is increased to hit the compression region and AM/PM represents as the change of the carrier
signal phase by an amplifier when input power is increased to hit the compression region [8].
AM/AM distortion happens while the amplifier set foot in the compression zone. In figure 1.3 the
1 dB compression point is shown. The high peaks of the input signal will drive the Power Amplifier
to hit the compression region and the gain of PA degrades from its constant value gain. A DPD
technique is a solution for this distorted amplitude and phase modulation problem.
Figure 1.5 AM/PM Conversion
The static and adaptive effects of the Digital Predistortion to amplitude and phase modulation of
the non- linear Power Amplifier will be discussed in the results section. In figure 1.6 AM/AM and
AM/PM curves show effect of the compression region when PA enters there.
17
Figure 1.6 AM/AM and AM/PM Curves [8]
1.5 Thesis Outline
The thesis is organized as follows. In chapter 2, discussed about power amplifier classes and
efficiency, chapter 3 includes 13.56MHz power amplifier design and 5 effects of Power Amplifier
nonlinearity. Chapter 4 have discussed about three different Linearization techniques.
In chapter 5, we discussed about the behavioral model of a Power Amplifier by using five different
methods, simulation methods and the method of DPD coefficient obtain.
In chapter 6 the digital predistortion modelling and equations are explained and discussed about
the inverse learning structure.
In chapter 7 we discussed about the model and the simulation results including static and adaptive
DPD. After this chapter, conclusions and future work for hardware implementation and wireless
communication systems are discussed.
18
INTRODUCTION TO POWER AMPLIFIER AND ITS
EFFICIENCY
2.1 Power Amplifier Classes and Efficiency
The most common and known as best class of power amplifier due to low distortion level is called
Class A amplifiers. Class A amplifier has highest linearity using single transistor process
technology like BJT, FET or IGBT etc. These transistors connect with a common emitter for the
both halves of the waveform, although it does not pass any base signal which represent that the
output is never driven into saturation region. However, base biasing Q point operate in the middle
of the line which makes the transistor never turns OFF.
Figure 2.1 Class A Amplifier Design and Output
Class B amplifier came up with a solution of efficiency and heating problem which creates in Class
A amplifier. However, Class B has higher efficiency, as one-half cycle is used as input, it creates
a higher distortion than Class A amplifier. So, the input power is not utilizing fully. To overcome
this problem, class B amplifier has been introduced.
The DC power is small here in class B as there is no DC biasing current which basically makes
the efficiency higher for Class B.
19
Figure 2.2 Class B Amplifier Design and Output
Class AB is one of the popular and common type of PA use in audio industry. Class AB is a
summation of Class A and Class B amplifier. A current mirror can be implemented by using a pair
of small signal diodes. Current mirror is a device whatever happens in one section of the transistor
it will reflect on the other section of the diode. Current mirror helps to eliminate the cross over
distortion because the bias voltage the diodes require it will go through those two transistors to
turn those ON. And it will be ON for more than a half cycle and creates the conduction angle
180oand 360o based on the biasing point.
Figure 2.3 Class AB Amplifier Design and Output
20
The Class C Amplifier represent the greatest efficiency but has the lowest operating cycle and
linearity. Although, due to heavily biased of class C amplifier, it stays on for less than half of an
input cycle and thus a conducting angle somewhere around vicinity of 90 degrees. Therefore, it
creates high but on the other hand it causes high distortion in the output signal. So, it does not meet
the criteria to use in audio industry.
Figure 2.4 Class C Amplifier Design and Output
They are using in certain RF application where efficiency is a key thing. They work in two
operating modes, one is tuned and another one is un-tuned, and they also have low power
dissipation. The output converts to sinusoidal wave by using LC resonance tank and connect with
collector of a transistor.
Power Amplifier designs in different high frequency relates with different switching techniques to
achieve the efficiency with less distortion. Some amplifiers consist RLC circuit with a combination
of Pulse Width Modulation (PWM) as an input and Digital signal processing (DSP) technique to
reduce the power loss. There are also switch mode power amplifiers which promises a great
efficiency with low distortion. Figure 2.5 and Table 1.1 shows the conduction angle, efficiency
and distortion level to provide a brief idea about all class of PA’s:
21
Figure 2.5 Amplifier’s Efficiency and Conduction Angle
Table 1.1 Amplifier’s Efficiency, Layout Area and Distortion
Amplifier Maximum Efficiency Layout Area Distortion
Class A 50% Small Low
Class B 78.50% Small High
Class AB <78.50% Small Moderate
Class C Approaching 100% Moderate Moderate
Class D Not Suitable for Narrowband RF
Class E Approaching 100% Moderate Low
22
POWER AMPLIFIER DESIGN AND IT’S NON-
LINEARITY
3.1 13.56MHz E- Class Power Amplifier Design and Simulation
In this section we have discussed about the analytical analysis of a Class-E PA, showing the
relation between the circuit component and the input parameters. The solution represents many
equations in analytical analysis section for the Class-E PA shown in Figure 3.1 due to value the
duty cycle switching and the RF configuration. We have used MATLAB Simulink to design this
Power Amplifier and later predistorted by using Simulink. We discussed about the capacitance
current, nonhomogeneous other nth number of order differential equations.
Figure 3.1 13.56 MHz E- Class Power Amplifier Schematic Design
3.1.1 Analytical Analysis
Two fundamental boundaries have used together to solve and relate the relation between the
parameters and element of the circuit with the E class Power Amplifier conditions in ((3.1) and
(3.2)) equations:
𝑉𝑐 (2𝜋
𝜔) = 0 (3.1)
𝑑𝑉𝑐(𝑡)
𝑑𝑡|𝑡=2𝜋/𝜔 = 0 (3.2)
The load current is IR(t) which should be sinusoidal wave. In theory, it will work if Q consist with
LC tank circuit in the schematic and this sine wave assumption can simplify the analysis part
23
𝐼𝑅(𝑡) = 𝐼𝑅 sin(𝜔𝑡 + 𝜑) (3.3)
If the switch is closed, the time interval 0 <t · π/ω, and the capacitance voltage Vc(t) = 0 and the
current through the capacitance (C) is, Ic (t) = C dVC (t) dt = 0 [1]. In this time interval, the switch
current IS(t) is
𝐼𝑆(𝑡) = 𝐼𝐿(𝑡) + 𝐼𝑅(𝑡)
=𝑉𝐷𝐷
𝐿𝑡 + 𝐼𝐿(0) + 𝐼𝑅𝑠𝑖𝑛(𝜔𝑡 + 𝜑) (3.4)
Where 𝐼𝐿(0) = −𝐼𝑅sin (𝜑)
When the switch is opened, the time interval d·π/ω < t < 2π/ω. Then, in the PA the current through
capacitance C is [1]
𝐼𝑐(𝑡) = 1
𝐿∫ (𝑉𝐷𝐷
1
𝑑𝜏/𝜔
− 𝑉𝑐(𝑡))𝑑𝑡 + 𝐼𝐿(𝑑𝜋
𝜔+ 𝐼𝑅(𝑡) (3.5)
Relation (3.5) can be re-arranged in the form of a linear, nonhomogeneous which has as solution
𝐿𝐶𝑑2𝑉𝑐(𝑡)
𝑑𝑡2+ 𝑉𝑐(𝑡) − 𝑉𝐷𝐷 − 𝜔𝐿𝐼𝑅 cos(𝜔𝑡 + 𝜑) = 0 (3.6)
𝑉𝑐(𝑡) = 𝐶 cos(𝑞𝜔𝑡) + 𝑉𝐷𝐷 −𝑞2
1 − 𝑞2𝑉𝐷𝐷 cos(𝜔𝑡 + 𝜑) (3.7)
Where, q= 1
𝜔√𝐿𝐶 (2.8)
p =ωL𝐼𝑅𝑉𝐷𝐷
(9)
𝐶 = {𝑞2 cos(2𝑞𝜋) cos(𝜑)
1 − 𝑞2p +
sin (2𝑞𝜋)𝑞𝑠𝑖𝑛(𝜑)
1 − 𝑞2p − cos (2qπ)}𝑉𝐷𝐷 (3.8)
The coefficients C came from the Class-E equations (3.1) and (3.2). It follows from (3.4) and (3.7)
that Vc (t) and IS(t) can be represent as VDD and ω only if d, q, p and ϕ are known.
3.1.2 Achievable Waveforms
This 13.56 MHz Power Amplifier waveforms shows different peak voltage and peak current values
though they have their own pros and cons values. Therefore, of Class-E waveforms provides a
significance from an application point. Designing class E 13.56MHz using a single transistor
where that transistor works as a switch, as in Figure 3.2, the following properties of Class-E
waveform plays significance roles:
24
• The peak value of switching voltage supposed to be too low but it should not cross the
breakdown voltage limit of the transistor.
• When the switch is turning OFF, on that moment transistor current should be low
specifically while transistor drives by a sine wave signal.
• To minimize the power losses, the rms value of the transistor current, the inductor (L) and
the capacitor (C) current, and the load current supposed to be low.
Figure 3.2 Class-E Waveforms With Ideal Components And Under Optimal Switching
Conditions
25
Figure 3.3 13.56MHz PA Output
26
3.2 Effects of Power Amplifier Nonlinearity
3.2.1 Harmonic Generation
Equation (3.1) is used to characterize the power amplifier behavior. Suppose we have a single tone
input signal which is representing by Vi. So, Vi (t) = cos(2πft). Based on Equation (3.1), applying
Vi (t) as input, the output V(t) has frequency components (f) along with nf where n represents the
integer [4]. This n integer number creates the nth harmonic in the power amplifier. These harmonics
occurs harmonic distortion, but they are not suppressed so it is not so complex technique to filter
those from the tone of the spectrum [2].
3.2.2 Intermodulation Distortion
Equation (3.1), also shows the input and output nonlinear relation. If there is a two-tone signal
input, Vi(t) = cos (2πf1 t) + cos (2πf2 t), going towards the amplifier, the output of the PA consists
some frequency as ±m f1±n f2. These are called intermodulation products. Given k = |m|+|n|, ±m
f1±n f2 products are called the kth intermodulation product of the power amplifier. If these
components are not suppressed, it causes distortion [2]. These components cannot be filtered out
due to the intermodulation frequency components are not too far away from the fundamental
components.
3.2.3 Spectral Regrowth
Spectral growth works like intermodulation distortion. Let’s assume we are going to pass a sine
wave signal through an amplifier and as an output low amplitude are unchanged and high
amplitude gets distorted. So, if we look at the frequency domain we might have some pulse shape
as an input but after passes through the nonlinear amplifier there are also some extra frequency
component and that is the spectral growth because of the nonlinear operation of the amplifier. It is
easier to find out if a single tone signal works as an input. But if any broadband signal passes
through the nonlinear channel it causes adjacent channel interference and the shoulder of the
spectrum will grow. Here creates an Adjacent Power Ratio (ACPR) which generally keeps low [2].
In today’s wireless communication systems this power ratio is vastly using to figure out the
linearity.
27
Figure 3.4 Spectral Regrowth
3.2.4 Cross Modulation
The nonlinear power amplifier channel works to transfer the signal of one carrier of frequency ω1,
to another carrier of frequency ω2 [4]. Assume that the input given in Equation (3.9) with carrier
of frequency ω1 which is modulated and carrier of frequency ω2 which is unmodulated is passing
through the PA
𝑉𝑖𝑛(𝑡) = 𝑉1 𝑐𝑜𝑠(𝜔1𝑡) + (1 + 𝑚(𝑡)) 𝑐𝑜𝑠(𝜔2𝑡) (3.9)
Here, m(t) represents modulating waveform.
𝑉𝑜𝑢𝑡(𝑡) = 𝛼1𝑉1𝑖𝑛(𝑡) + 𝛼2𝑉2 𝑖𝑛(𝑡) + 𝛼3𝑉3 𝑖𝑛(𝑡) (3.10)
This is the PA model up to 3rd order output where the output equation of the power amplifier
consists component of ω2 as α1V1 + α3V 3 + α3V1(1 + m(t))2 cos(ω1t) which represents frequency
ω2 as unmodulated signal of the input is transferring ω1 which is modulated within the same
bandwidth after passing through the nonlinear RF system. This phenomenon is called cross
modulation.
28
3.2.5 Desensitization
Desensitization is an EMI form where receiver is not able to receive some weak signal and
blocking by some unwanted strong signal which works in different frequency, however it seems
to be able to receive while in the non-linear system there is no interface [4].
So, the input would be;
Vi (t) = V1 cos(ω1t) + V2 cos(ω2t) (3.11)
Here the desired signal is ω1, the unwanted signal is ω2 and V2 V1. If we use equation (3.10), to
obtain output, then the output signal has the wanted signal of frequency ω1 as; α1V1 + α3V3+
α3V1V2 2 cos(ω1t). Here, α3 is gives the PA saturation as it drives in negative. So that means ω2 is
dominating the wanted signal ω1. This phenomenon is known as desensitization.
29
LINEARIZATION TECHNIQUES
4.1 Linearization Techniques
Now after looking at the effect of non-linearity, in this section we discussed about the removal
techniques of power amplifier’s nonlinearity. Those transmitter linearization techniques are:
• Feedback
• Feedforward
• Predistortion
In this section we will discuss about the predistortion techniques in detail.
4.1.1 Feedback
Feedback is a linearization technique can be applied either RF or Baseband signal. The block
diagram has shown in Figure 4.1. Here, Vi(t) is the input signal, Vo(t) is the output signal. The
output signal is feedback with a β gain and Vr(t) signal is obtained. The difference of Vi(t) and
Vr(t) gives the error signal Ve(t). The error signal is given to the amplifier and linear output is
obtained.
Figure 4.1 Feedback Linearization [4]
The main advantage of feedback linearization technique is that the closed loop gain gets less
sensitive to the gain variation of the power amplifier. Furthermore, it is easy to add the 16-
necessary circuitry to the power amplifier for feedback [4]. The main disadvantage is the delay
between the input and copied output signal which causes a limitation of gain and bandwidth
product. This affects the stability issue.
30
4.1.2 Feedforward
The block diagram of Feedforward method has shown in Figure 4.2. This is a technique which
contains two power amplifiers and one of the power amplifier need to linearize. So, input signal
divide into two signals. One goes for main power amplifier where linearization is a goal and the
second input signal works to estimate the error. The loop calls nulling loop and the second loop
known as error cancellation loop. Main Amplifier has the highest power and the error amplifier
(Figure 4.2) helps to linearize the main amplifier which has smaller gain and does not work into
much higher power level. In the nulling loop, this technique removes the original tone and only
inter modulated distortion remains for the error amplifier.
Figure 4.2 Feedforward Linearization [4]
The advantages of this technique are the gain of the main power amplifier does not reduce. This
system is very stable align with higher bandwidth. The disadvantage is system is not adaptive
which means characteristics of PA does not change with temperature, ageing or heat changes in
the circuitry. And bandwidth changes rapidly with respect to power amplifier’s input and output
matching circuit.
4.1.3 Predistortion
Predistortion is a technique that used to improve the linearity of radio transmitter amplifier. The
predistortion circuit inversely model the amplifiers gain and phase characteristics. In a sense,
inverse model introduced in the input of the amplifier so that they could cancel all the non-linearity
31
that amplifier might have [4]. It is a cost saving and power efficiency technique. It can be
implemented in both analog and digital manner known as digital predistortion. Analog
Predistortion uses non-linear devices such as anti- parallel diodes.
Figure 4.3 Predistortion Method of Linearization [4]
The three advantages of Predistortion technique is:
• Wideband linearization.
• Stable.
• Conceptually simple.
The predistortion can be categorized in three parts according to the frequency at which it is
implemented;
• RF predistortion
• IF Predistortion
• Baseband Predistortion
RF predistortion works as an analog predistortion and compensate with AM/ AM and AM/PM
modulation, Harmonics effect, memory effect and linearize PA output behavior as DPD. The IF
predistortion means intermediate frequency which predistorter to works in various frequency level.
Therefore, IF works more adaptively over the RF predistortion technique which creates a narrower
bandwidth in spectrum. IF are used in radio receiver which has an incoming signal shifted for an
32
amplification before the final detection is done. It is easier to tune and to make sharply selected
filter lower frequency at lower fixed frequency. Using a very few components can save money
with no stability issues are grown.
Figure 4.4 a) RF predistortion, b) IF predistortion, c) Baseband predistortion [4]
One of the disadvantage is about the linearity. At high frequency distortion is still there, and if in
case transfer function changes, fabrication is needed for the circuity which indicates that adaptively
will not work for power amplifiers.
Therefore, Baseband Predistortion is more adaptable and easier predistortion technique than others.
In the following section we discussed about the Baseband Predistortion technique in detail. There
are two types of predistortion based on the adaptiveness [4];
• Adaptive Predistortion
• Non-adaptive Predistortion
Changing temperature with time, antenna matching output could affect amplifier characteristics.
Finally, the predistortion can be categorized in terms of handling the memory effects of the power
amplifier; and also, the baseband predistortion modelling is discussed in Chapter 5.
• Memoryless Predistortion
• Predistortion with Memory
33
BASEBAND PREDISTORTION MODELLING OF NON-
LINEAR SYSTEMS
5.1 Introduction
This chapter deals with the problem of modeling nonlinear passband systems by the means of
quasi-memoryless models and Volterra series-based models in the complex baseband domain. The
resulting baseband models can be employed, e.g., in simulations, to assume the created distortion,
without considering a frequency up-conversion unit which shifts the baseband signal to the RF
carrier frequency [9]. Since with these baseband models, we only relate the complex input and
output envelopes, the computational complexity to calculate the output signal of the nonlinear
model can be significantly reduced.
In Sec. 5.1 we introduce complex nonlinear baseband modeling and review existing literature. In
Sec. 5.2 we transform a static nonlinear system which is composed of two polynomial functions
acting on two orthogonal carriers and a linear passband filter to the baseband domain. We either
obtain a so called memoryless or a quasi-memoryless nonlinear model, depending on whether the
resulting parameters are real or complex. In Sec. 5.3 we replace the passband nonlinearity from
Sec. 5.2 by a real Volterra series. In Sec. 5.4 we consider the relationship between a quasi-
memoryless baseband model and a complex Volterra series model. Amplitude and Phase
modulation concept could extend for the case of a Volterra model in Sec. 5.5 and use the frequency
domain-based AM/AM and AM/PM surfaces to identify the complex linear filters of a memory-
polynomial model.
5.2 Bandpass models versus baseband models
Power Amplifier used in wireless communication systems can be described as nonlinear functions
that map a real-valued RF signal or band-pass signal to a real valued RF output. Although, a PA
model can be constructed using the RF (or bandpass) input-output observations, by assuming that
the PA input signal is narrowband, the modeling of PAs can be greatly simplified. It is shown that
the information carried by a band-pass RF signal URF(t) centered around a frequency fc can be
34
completely represented using its complex-valued baseband equivalent. The relation between a
narrowband band-pass signal URF(t) and its complex baseband signal, u(t) is given by
𝑢𝑅𝐹(t) = A(t) cos ( 𝜔𝑐t + φ(t))
= Re{A(t)𝑒 j( 𝜔𝑐t + φ(t)) }
= Re{u(t)𝑒jφ(t) } (5.1)
with u(t) = A(t)e jφ(t). A(t) represents amplitude and φ(t) represents phase modulation, and ωc
denotes the RF carrier angular frequency. Using this relation, PA bandpass models can be
translated into the baseband domain. To understand this, we presented a simple example of the
modeling of a PA. If the nonlinear behavior of a PA can be approximated using a polynomial
model given by
𝑦𝑅𝐹 = ∑𝑎P
P
P=0
𝑢𝑅𝐹P (𝑡)(5.2)
where ap = parameters of the model and P = maximum nonlinear order. Assume that the input
signal is the bandpass signal which is URF(t). To make it simple, assume that the PA can be
modeled with a 3rd order polynomial, i.e. P = 3
𝑦𝑅𝐹(𝑡) = 𝑎0 +𝑎2𝐴
2(𝑡)
2+ [𝑎1𝐴(𝑡) +
3𝑎34𝐴3(𝑡)] cos( 𝜔𝑐t + φ(t))
+𝑎2𝐴
2(𝑡)
2cos(2 𝜔𝑐t + 2 φ(t)) +
𝑎3𝐴3(𝑡)
4cos(3 𝜔𝑐t + 3 φ(t)) (5.3)
Note in (5.3) that besides the signal at the carrier frequency ωc, three new signals are generated: a
DC component and two signals at frequencies 2ωc and 3ωc, also known as 2nd and 3rd order
harmonics, respectively. The DC component is generally blocked by a capacitor at the PA output
matching network and the harmonics are filtered out at the PA output. Therefore, the only signal
close to the carrier frequency is given
𝑦𝑅𝐹,𝜔𝑐(𝑡) = [𝑎1𝐴(𝑡) +3𝑎34𝐴3(𝑡)] cos( 𝜔𝑐t + φ(t)) (5.4)
Considering that in wireless systems, the carrier frequencies are in the order of GHz and the
bandwidth of the signal in the order of MHz, even after bandwidth expansion, yRF,ωc(t) can also
be narrowband. Therefore, it can be represented in the baseband domain as,
𝑦(𝑡) = 𝑎1𝐴(𝑡)𝑒𝑗𝜑(𝑡) +
3𝑎34𝐴(𝑡)3𝑒𝑗𝜑(𝑡)
35
= 𝑎1𝑢(𝑡) +3𝑎34𝑢(𝑡)|𝑢(𝑡)|2(5.5)
Where y(t) is the baseband equivalent of the RF output signal and |.| denotes the absolute value.
The baseband input and output relation can be generalized to an arbitrary nonlinear order to create
the baseband model
𝑦(𝑡) = ∑𝑏P
P
P=0
𝑢(𝑡)|𝑢(𝑡)|(P−1)(5.6)
Where bp denotes the parameters of the model. This model is known as the baseband polynomial
model and is one of the simplest models used to characterize a PA. From (5.6), it can be observed
that in the baseband representation of the polynomial model only odd-order terms are present. The
distortions caused by the even order components are far from the carrier frequency and do not
donate to the baseband output y(t). However, it was shown that better accuracy can be obtained if
even-order terms are also considered. In this chapter, both even and odd order terms have used in
the polynomial based models. Characterizing the PAs in the baseband domain reduces the
computational complexity, since the input and output signals can be acquired at lower sampling
rates. Although bandpass behavioral models for PAs have been proposed in the literature, most of
the behavioral models published are constructed in the baseband domain. Therefore, all the models
considered in this chapter are of the baseband type.
5.3 The Quasi-Memoryless Case
Memory-Polynomial Model with Quasi-Memoryless Models as Saleh Model:
Quasi-memoryless models consider both AM/AM and AM/PM transfer functions. Static DPD
design output provides the accuracy when the memory effects are not essentials. For narrow band
application, quasi memoryless shows better output precision.
The Saleh model:
The Saleh model is a quasi-memoryless model which used by four parameters [𝛼𝑎 , 𝛽𝑎, 𝛼𝜑 , 𝛽𝜑 ] to
make the suitable model [10]. It is amplitude modulation and phase modulation transferred
function that are described by these equations:
𝑔(𝑟(𝑛)) =𝛼𝑎𝑟(𝑛)
1 + 𝛽𝑎𝑟(𝑛)2
36
𝑔(𝑟(𝑛)) =𝛼𝜑𝑟(𝑛)
1 + 𝛽𝜑𝑟(𝑛)2
Where, [𝛼𝑎, 𝛽𝑎, 𝛼𝜑, 𝛽𝜑 ] are the model's parameters.
5.3.1 AM/AM- and AM/PM-Conversion
To see that the quasi-memoryless model can be fully represented by the AM/AM conversion and
AM/PM-conversion (cf. Figure 2.2), we expand with x˜(t) = a(t) exp (jφo(t)), which results in [9]
𝑦(𝑡) = 𝐻[𝑥(𝑡)]
= 𝑒𝑥𝑝(𝑗𝜑𝑜(𝑡)) ∑ 𝑑2𝑘+1[𝑎(𝑡)]2𝑘+1
[𝐿2]−1
𝑘=0
= |𝑣(𝑎(𝑡))| exp (𝑗(𝜑0(𝑡) + arg{𝑣(𝑎(𝑡))})) , (5.7)
Where the complex function,
𝑣(𝑎(𝑡)) = ∑ 𝑑2𝑘+1[𝑎(𝑡)]2𝑘+1
[𝐿2]−1
𝑘=0
(5.8)
depends purely on the magnitude a of the complex input signal x˜(t) [9]. The function |v(a)| in (5.7)
describes the AM/AM-conversion and the function arg {v(a)} in (5.7) describes the AM/PM-
conversion. These nonlinear functions are calculated by
|𝑣(𝑎)| = √[∑ ∑ 𝑑2𝑘+1𝑑2𝑙+1𝑎2(𝑘+𝑙+1)
[𝐿2−1]
𝑙=0
[𝐿2−1]
𝑘=0
] (5.9)
And
arg {v(a)} = arctan{∑ 𝑙𝑚 {𝑑2𝑘+1}𝑎
2𝑘+1[𝐿2]−1
𝑘=0
∑ 𝑅𝑒 {𝑑2𝑘+1}𝑎2𝑘+1[𝐿2]−1
𝑘=0
} (5.10)
respectively. Figure 5.7 depicts the quasi-memoryless complex baseband model H which is
equivalent, (y(t) = Re{y(t)e jωct})
37
5.3.2 Frequency-Domain Representation
To express the output signal of the quasi-memoryless system y(t) in (2.16) in the frequency domain,
we apply the Fourier transform denoted by F to (2.16), which yields
Y (ω) = F {y˜(t)}
= ∑𝑑2𝑘+1(2𝜋)2𝑘
[𝐿2−1]
𝑘=0
𝑋(𝜔) ⋆ …𝑋(𝜔) ⋆ X ∗ (−ω) ⋆ …⋆ X ∗ (−ω), (5.11)
where X(ω) = F {x˜(t)}, and ⋆ denotes the convolution operator.
If we consider stationary stochastic signals, we cannot calculate the Fourier transform because of
their infinite energy. In this case, the spectral characteristics of the output signal y˜(t) is obtained
by computing the Fourier transform of the auto-covariance function of y˜(t) in terms of its power
spectrum density [9].
Figure 5.1 Baseband Model of a Quasi-memoryless System, which is Composed of Two Static
Non-linear Functions Described By The AM/AM Conversion |v(a)| and AM/PM Conversion arg
{v(a)}. [9]
5.4 Complex Baseband Modeling with Volterra Series
To overcome the problem of generating an asymmetric power spectrum at the output of a quasi-
memoryless model (if the magnitude of the input signal spectrum is symmetric), we must introduce
some memory into the complex baseband model. Volterra series is a mathematical tool which can
describe weak nonlinear systems with memory effects.
38
5.4.1 Time-Domain Representation
We replace the nonlinear passband operator G, which is composed of two polynomial series and
a Hilbert transformer, by a more general operator G whose functional description is given by the
Volterra series
u(t) = G[ x(t)] = ∑𝑢𝑙
𝐿
𝑙=1
(𝑡)
𝑢𝑙(𝑡) = ∫ …∫ ℎ𝑙(𝜏1… , 𝜏𝑛)∏𝑥(𝑡 − 𝜏𝑖)𝑑𝜏𝑖
𝑙
𝑖=1
∞
0
∞
0
(5.12)
Where, hl = lth-order time-domain Volterra kernel and L = highest order of the real passband
nonlinearity
The output signal is filtered by a zonal filter which has the linear operator F to control the unwanted
signals located around the actual carrier frequency ±ωc [9]. Therefore, the output signal of the
linear filter
y(t) = ( F ° G) [ x(t)]
= 𝐻[𝑥(𝑡)] = ∑𝐹[𝑢𝑙
𝐿
𝑙=1
(𝑡)](5.13)
includes only the spectral signal which are around the carrier frequency ±ωc [9]. To express the
lth-order term of the output signal of the passband Volterra system ul(t) in (5.12), the product in
(5.12) is expressed with the time delayed version of the passband input signals in a mathematical
closed form by
∏𝑥(𝑡 − 𝜏𝑖
𝑙
𝑖=1
) =1
2𝑙∑ …
2
𝑘𝑙=1
∑(∏𝑥𝑘𝑖
𝑙
𝑖=1
2
𝑘𝑙=1
(𝑡 − 𝜏𝑖) exp(𝑗𝜔𝑐∑(−1)𝑘𝑖
𝑙
𝑖=1
𝜏𝑖)
× exp(𝑗𝜔𝑐 𝑡∑(−1)𝑘𝑖+1𝑙
𝑖=1
(5.14)
where the signals x1(t) = ˜x(t) and x2(t) = ˜x ∗ (t) in (5.14) are introduced for a convenient
representation. The lth-order output signal of the passband Volterra system in (5.12) is expressed
with (5.12) by
39
𝑢(𝑡) =1
2𝑙∑ …
2
𝑘𝑖=1
∑∫ …∞
0
2
𝑘𝑖=1
∫ ℎ𝑙
∞
0
(𝜏1, … , 𝜏𝑙) (∏𝑥𝑘𝑖
𝑙
𝑖=1
(𝑡 − 𝜏𝑖))
× exp (𝑗𝜔𝑐∑(−1)𝑘𝑖
𝑙
𝑖=1
𝜏𝑖)𝑑𝜏1…𝑑𝜏𝑙 × exp (𝑗𝜔𝑐𝑡∑(−1)𝑘𝑖+1𝑙
𝑖=1
)(5.15)
Where, the product in (5.15) is composed of the permutations of the baseband signal x1(t) = x˜(t)
and its conjugate x2(t) = ˜x ∗ (t), respectively. The overall output signal y(t) in Figure 2.4 is
calculated with (5.13) by applying the linear operator F to the lth-order output signals of the
passband Volterra system in (5.15) for l = 1, . . ., L. Each of the 2l l-fold convolution integrals in
(5.15) which contributes to the lth order output signal ul(t) is multiplied by a phasor which
corresponds to integer multiples of the carrier frequency. Therefore, the spectra of the convolution
integrals are shifted in the frequency-domain to the corresponding multiples of the carrier
frequency 𝜔𝑐𝑡 ∑ (−1)𝑘𝑖+1𝑙𝑖=1 . If the order of the nonlinearity l ∈ No, the spectra can only be around
the odd multiples of the carrier frequency up to lωc. If the order of the nonlinearity l ∈ Ne, the
spectra can only be located around the even multiples of the carrier up to lωc. The bandwidths of
the individual contributions to (5.15) are 2 l-times the bandwidth B of the complex baseband signal
x˜(t). If the carrier frequency satisfies ωc ≥ B (2L − 1) the spectra of the individual contributions
from (5.15) remain separate [9]. Therefore, the output signal passed by the linear 1st-zonal filter
can be calculated from (5.15) if the carrier phasor in (5.15) is constraint to be
exp (𝑗𝜔𝑐𝑡∑(−1)𝑘𝑖+1𝑙
𝑖=1
) = exp (±𝑗𝜔𝑐𝑡)(5.16)
This equality can only be satisfied for the odd orders of the nonlinearity l ∈ No. Therefore, solely
the odd orders of the nonlinearity contribute to the filtered output signal which surround the carrier
frequency ±ωc. For the even orders of the nonlinearity l ∈ Ne, the filtered output signal F[ul(t)] =
0. From [9] contributions in (5.15) only 2 (𝑙𝑙−1
2
)terms fulfill (5.16). Without any loss of generality,
the passband Volterra kernels hl in (5.15) are assumed to be symmetric and, therefore, the
permutations of the product in (5.15) are identical for (𝑙𝑙−1
2
)terms. The same is obtained for the
second group of terms which are the conjugate of the first one. Therefore, the filtered output signal
is given by
40
𝐹[𝑢𝑙(𝑡)] =1
2𝑙(𝑙
𝑙 − 1
2
)(𝑓𝑙(𝑡) + 𝑓𝑙 ∗ (𝑡)
= 2𝑅𝑒{1
2𝑙(𝑙
𝑙 − 1
2
) (𝑓𝑙(𝑡)}(5.17)
where the function f1 in (5.17) is expressed with l ∈ No by
= ∫ …
∞
0
∫ ℎ𝑙
∞
0
(𝜏1, … , 𝜏𝑙) exp
(
−𝑗𝜔𝑐
(
∑𝜏𝑖
𝑙+12
𝑖=1
− ∑ 𝜏𝑖
𝑙
𝑖=𝑙+32 )
)
×∏𝑥1
𝑙+12
𝑖=1
(𝑡 − 𝜏𝑖) ∏ 𝑥2
𝑙
𝑖=𝑙+32
(𝑡 − 𝜏𝑖)𝑑𝜏1…𝑑𝜏1 exp(𝑗𝜔𝑐𝑡) (5.18)
The final output signal passed by the linear 1st-zonal filter is given with (5.13) by
𝑦(𝑡) = ∑ 𝐹[𝑢2𝑘+1
[𝐿2]−1
𝑘=0
(𝑡)]
= 𝑅𝑒{∑ ∫ …∫ ℎ2𝑘+1(𝜏1, … , 𝜏2𝑘+1)∏𝑥(𝑡
𝑘+1
𝑖=1
∞
0
∞
0
[𝐿2]−1
𝑘=0
− 𝜏𝑖) ∏ 𝑥∗(𝑡 − 𝜏𝑖) × 𝑑𝜏1…𝑑𝜏2𝑘+1exp (𝑗𝜔𝑐𝑡)}
2𝑘+1
𝑖=𝑘+2
(5.19)
where the variable substitution k = (l − 1)/2 is introduced for a more convenient representation of
(5.19). The baseband-equivalent Volterra kernels in (5.19) are defined with (5.18) by
ℎ2𝑘+1(𝑡1, … , 𝑡2𝑘+1) = 1
22𝑘(2𝑘 + 1𝑘
)ℎ2𝑘+1(𝑡1, … , 𝑡2𝑘+1) × exp (−𝑗𝜔𝑐(∑ 𝑡𝑖
𝑘+1
𝑖=1
− ∑ 𝑡𝑖))
2𝑘+1
𝑖=𝑘+2
(5.20)
The term baseband-equivalent means that the frequency-domain representation of the kernels in
(5.20) contains some frequency components around the zero frequency (baseband) and −2ωc. The
latter one does not contribute to the output signal y(t) in (5.19), because the frequency-domain
representations of the baseband signals x˜(t) and x˜ ∗ (t) are zero around −2ωc.
Except for the carrier phasor exp (jωc t), the signal within the braces of (5.19) represents the
nonlinear passband output signal y˜(t) in the baseband domain because y(t) =Re {𝐲 ˜(𝐭) exp(𝐣𝛚c𝐭)}.
41
Figure 5.2 Complex Baseband Time-Domain Output Signals of a 13.56 MHz RF PA, a
Memoryless PA Model.
5.5 Comparison of the Model Structure Performances of the Predistorters
In order to compare different predistortion technique performance, the linearity parameters are
used in the spectrum analyzer for frequency domain representation and the error vector magnitude
(EVM) in the time domain representation. Moreover, the model’s complexity is an essential
component to compare as the predistortion are performed in real system. So, the linearity and
model’s complexity are used for comparison. As we mentioned before about the memoryless
predistortion and memory predistortion, it is categorized to work on the dynamic non-linearity.
The DPD algorithm does not take care of memory effects which degrade the output signal as the
increase of the input signal. So, look up table (LUT) method takes into account of memory effects
to provide a promising linear performance. In the memory predistorters: The Volterra series and
memory polynomial predistorters provide better linearization behavior than the others. Their
performance is better because Volterra series is a complete model than the others which are arisen
by Volterra model but provides distorted performence. However, Volterra series and memory
42
polynomial have promising performance on linearity, Volterra series is a bit more complex model
which makes a con of this model. Static non-linearity dominates the parameter estimations of the
nonlinearities. The memory effect nonlinearity is located separately in two nonlinear boxes.
Additionally, two nonlinear box models have better performance than the Wiener and
Hammerstein models where both models are static and dynamic nonlinear. But two nonlinear class
models the dynamic nonlinearity with a memory polynomial function described with a first and
second order nonlinearity. So, two nonlinear box models are a better performance model, but it is
complex. Therefore, memory polynomial predistortion technique is a most constructive with
favorable output method considering linearization and complexity. So, in the following section,
memory polynomial predistortion will be discussed in detail.
43
MEMORY POLYNOMIAL DIGITAL PREDISTORTION
6.1 Theory
In a memory polynomial predistotion techniques, to make up with the PA non-linearity behavior
the input signal gets distorted. Baseband is a method that’s how the DPD algorithm is applied here.
Figure 6.1 shows the block diagram of the DPD scheme:
Figure 6.1 Block Diagram of Digital Predistortion Scheme [4]
In this scheme, Baseband input is un and xn is an output of the predistoter as we can see in figure
6.1. xn input also works as an up conversion (mixer and filters) for the PA. The output of the PA
goes to antenna and some of them are down converted (mixers and filters) to normalize the small
signal gain and output yn is achieved. We used xn and yn as algorithm. Here, in the predistortion
portion we have used predistotion function and this scheme divided into two parts. The first part
is to characterize the PA model accurately which is call forward modelling [4]. Therefore, here PA
is known as forward model. Then the inverse function comes to the picture and predistortion
function can be achieved. So, the predistortion function works as an inverse model which makes
the complexity for the predistortion to find out the PA parameters.
Here, the PA and predistortion is known as memory polynomial because both are modeled with
same function. So, the memory polynomial can describe as Volterra form which is one of the
advanced model where the diagonal terms are under discussion which decreased the number of
co-efficient. The memory polynomial is defined as in Equation (6.1) [4].
𝑧(𝑛) = ∑∑𝑎𝑘𝑞𝑥(𝑛 − 𝑞)|𝑥(𝑛 − 𝑞)|𝑘−1
𝑄
𝑞=0
𝐾
𝑘=1
(6.1)
Where, xn= the baseband complex input signal,
44
z(n)= the baseband complex output signal,
K= the polynomial (nonlinearity) order, and
Q= the memory depth.
Based on Equation 6.1, the output is not obtained with the input. It also based on the previous input
which defines that the memory effects are also in consideration. Moreover, the even order
nonlinear terms are included in modeling which can make the power amplifier linearization better.
Besides, x(n-q) term consists phase information where the input term is not phase blind, which
makes the model complete. While both amplifier and predistorter is modelling, the input and the
output of device under test is familiar. The trouble occurs to obtain the coefficients akq where their
characteristics are determined by finding these coefficients [4]. As we discussed before, the
amplifier is model works as a forward model while it is associated with memory polynomial. The
predistorter is known as an inverse model and how these inverse modelling works, described in
the section 6.2.
6.2 Inverse Structures
There are three methods of architecture of inverse model structure:
• pth-order inverse method
• Indirect Learning Architecture (ILA)
• Direct Learning Architecture (DLA)
6.2.1 pth-order Inverse Method
This method is an inverse is a technique used to find the inverse of nonlinear systems whose input-
output relation can be described by the Volterra series model. The p-th order inverse of a nonlinear
system is represents as p-th order Volterra series model that while connected with the nonlinear
system, results in a Volterra series model that is linear up to p-th nonlinear order. The p-th order
inverse has some drawbacks. Its derivation is computationally heavy, and it does not consider
terms higher than the p-th nonlinear order. Additionally, it represents stability problems when the
linear part of the system is not stable so that the p-th order inverse is not a popular technique to
estimate the predistorter parameters. However, the p-th order inverse theory developed in is
45
sometimes used to support the operation principle of the indirect learning architecture which will
be discussed in the next section [4].
6.2.2 Indirect Learning Architecture
One of the advantage of indirect pre-distortion learning architecture is – simple structure and low
complexity. Mainly, there is no need to know about the specific model and parameters of PA when
estimating coefficients of pre-distorter. Figure 6.2 shows the indirect pre-distortion adaptive
learning architecture where with the adjustment of gain, the output signal of amplifier z(n) is taken
as input signal of pre-distortion learning and training network. By comparison between pre-
distortion signal y(n) and output signal of training network �̂�(𝑛), error e(n) = y(n)- 𝑦^ (n) is
obtained as parameters which will be sent to the pre-distorter. If the algorithm converges, then e(n)
= 0, namely y(n) = 𝑦^(n) and z(n)/G= x(n). The indirect learning architecture introduces post
inverse structure which reduces the sensitivity of signal parameters and real-time closed-loop
system.
Figure 6.2 The Indirect Learning Architecture (ILA) For Non-linear Compensation.
6.2.3 Direct learning architecture
Another popular technique used to find out the parameters is the direct learning architecture (DLA),
illustrated in Figure 6.3 DLA is characterized for directly minimizing the error between the output
signal yd(n) and the actual output from the amplifier y(n), that is e(n) = yd(n) − y(n) Direct Learning
Architecture (DLA) uses complex optimization algorithms techniques to calculate the parameters
46
of the predistorter. Various DLA algorithms have been proposed in the literature, unfortunately
they are compound in structure, computationally expensive and slow convergence.
Figure 6.3 Direct Learning Architecture
DLA is generally implemented in two steps. First, power amplifier is a forward model. Once the
model is obtained, a nonlinear algorithm is used to estimate, through iterative processing, the
predistorter parameters that reduce the error between the output and the PA model output. When
the nonlinear algorithm gets solved, the calculated parameters leads to generate a predistorted
signal u(n) which works as an input for a real amplifier. This process is will continue until the
predistorter PA model reaches to the best solution.
6.3 The Improved Pre-Distortion Algorithm for Indirect Learning Architecture (ILA)
In accordance with the changes of estimation error, LMS adaptive algorithm adjusts tap
coefficients of a finite response filter automatically, so that the cost function could be minimized.
Suppose x (n) denotes the input vector of FIR filter whereas y(n) denotes the output vector of filter,
and tap coefficient of filter is represented by 𝜔(𝑛). Thus, 𝑦(𝑛) = 𝜔𝐻(𝑛)𝑥(𝑛).
Assume d(n) indicates the desired signal, then the difference e(n) between output signal of FIR
filter y(n) and desired signal d(n)could be denoted by:
𝑒(𝑛) = 𝑑(𝑛) − 𝑦(𝑛) = 𝑑(𝑛) − 𝜔𝐻(𝑛)𝑥(𝑛) = 𝑑(𝑛) − 𝑥𝑇(𝑛)𝜔(𝑛)(6.2)
The most common criterion in filter design is minimum mean square error criterion, in order to
minimize the mean square error between expected response and actual output of filter, we define
the cost function as below:
𝐽(𝑛)𝑑𝑒𝑓 = 𝐸{|𝑒(𝑛)|2} = 𝐸{𝑒|𝑑(𝑛) − 𝜔𝐻𝑥(𝑛)|2}(6.3)
47
The aim of filter design is to find the appropriate filter coefficient matrix 𝜔(𝑛) so that J(n) can
achieve its minimum. Solving by gradient descent method, here after:
𝜔(𝑛) = 𝜔(𝑛 − 1) −1
2𝜇(𝑛)∇𝐽(𝑛 − 1) (6.4)
Where,
∇𝐽(𝑛) = −2𝐸{𝑢(𝑛)𝑑∗(𝑛)} + 2𝐸{𝑢(𝑛)𝑢𝐻(𝑛)}𝜔(𝑛)(6.5)
The real gradient vector and instantaneous gradient vector are a pair of unbiased estimators:
𝐸{∇̂𝐽(𝑛)} = −2𝐸{𝑢(𝑛)[𝑑∗(𝑛) − 𝑢𝐻(𝑛)𝜔(𝑛 − 1)]} = ∇𝐽(𝑛) (6.6)
Replace ∇𝐽(𝑛 − 1)by instantaneous gradient vector ∇̂𝐽(𝑛 − 1),then:
𝜔(𝑛) = 𝜔(𝑛 − 1) + 𝜇(𝑛)e(n)x(n) (6.7)
If 𝜇(𝑛) is a constant, it is called fixed step LMS algorithm. Otherwise, we call it variable step LMS
algorithm.
LMS adaptive algorithm has been widely used in applied technology. Moreover, it has simple
calculation and meets well compensation performance. On the other hand, it is critical to select
step factor. If it sets as a small value, then the nonlinear distortion could not be well compensated.
However, if it is too large, the stability of system could not be guaranteed. Since the constellation
is merely a qualitative evaluation criterion, error vector magnitude EVM which is used to describe
PA pre-distortion performance, is a remarkable quantitative index to measure amplitude and phase
errors of modulated signal.
6.4 Estimation Algorithms
6.4.1 Least Mean-Square Algorithm (LMS)
Least Mean Algorithm works in way so that storage requirements will decrease to solve the matrix
solving requirements. When it needs to calculate the coefficients, the error feedback comes to the
picture. LMS is an easier model of algorithm which makes it efficient to apply. On the other hand,
LMS algorithm shows much slow convergence due to the high order nonlinearity coefficients are
very small. So, this algorithm is not realistic [4].
48
6.4.2 Recursive Least-Squares Algorithm (RLS)
Recursive Least-Squares Algorithm which consists low storage requirements relative to the matrix
solution. It modifies the inverse covariance matrix as a new sample which arrives by using a co-
efficient update. RLS covariance does not get affect by very small coefficient with fast converges
which makes biggest advantages of using RLS over LMS. But, RLS algorithm is much complex
than LMS [4].
49
SIMULATION RESULTS
7.1 Introduction
In this chapter, a workflow for DPD and power amplifier modeling are presented. We showed the
DPD block design and algorithm implementation from an off-line batch which is taking care of
the derivation including inverse matrix model to a fully adaptive application where no inverse
matrix is involved. We used spectral growth to evaluate the adaptive DPD design’s effectiveness.
We used a combination of MATLAB and Simulink for this project. MATLAB was used to define
parameters, individual algorithms to find out the coefficient output. Simulink has used to designing
the model and integrate the DPD and power amplifier together to show the spectral growth through
the feedback loop and data sizes.
7.2 Modeling and Simulating the Power Amplifier (Saleh Model)
The PA model contains a Saleh model 13.56 MHz power amplifier in series with a complex filter
as shown in Figures 7.1 and 7.2. the excitation is used here is called an Additive White Gaussian
Noise (AWGN) signal that will reduce the distortion by passing through a low-pass filter [6]. We
run the model and log the input and output signals, x and y, respectively when observing their
output in spectrum shown in Figure 7.3.
Figure 7.1 The E-class PA Measuring Model
50
Figure 7.2 The Structure of The Saleh Memoryless Non-linearity [6]
Figure 7.3 The PA Output Shows 26.49 dB of Gain at The Expense of Significant Spectral
Regrowth
7.3 Deriving DPD Coefficient:
We derived static DPD design from 13.56 MHz class E power amplifier calculation. Here, Figure
7.3 describe the illustrate derivation graphically. The top path in Figure 7.3 represents our power
amplifier model in Figure 7.1. the power amplifier is divided into a non-linear function which
comes after by a linear gain G. The middle path in figure 7.4 shows the power amplifier running
51
in reverse direction which shows the actual DPD. However, technically we can’t run power
amplifier automatically in reverse way, it is possible in analytical way and here DPD comes to
play the potential role. For the reverse way, we applied inverse nonlinear model, f-1(X1, X2, …,
Xn). The bottom path in Figure 7.4 is a cascade connection of the top two paths to get the linear
output [6].
Figure 7.4 DPD Derivation and Implementation [6]
7.3.1 Memory-Polynomial Model implementation using Matlab Simulink
A PA memory effect defines that the output of the PA is not only obtained by the current input but
also depends on the input that provided in previous. Memory effect is also defined by the change
of amplitude and phase modulation’s distortion to change in inter-modulation frequency.
The power amplifier nonlinearity is classified in two ways: static nonlinearity and dynamic
nonlinearity. The static nonlinearity defines as the amplitude and phase modulation graphical
representation of the current input of the PA which contains a well-built nonlinearity. The dynamic
nonlinearity communicates with the memory effect which contains a weak nonlinearity and
representable by a non-linear filter. Additionally, the dynamic nonlinearity also classified in two
types: one is linear memory effects and another one is the nonlinear memory effects [4].
52
Due to not ideal frequency response, the linear memory effect happens which represents by a filter
which is linear. On the other hand, due to impedance matching condition, biasing circuit modeling,
and harmonic frequencies the non-linear memory effect happens.
Figure 7.5 Basics of Memory Effect [4]
The memory effect is the summation of linear and non-linear memory effect and can be represent
with a filter which is non-linear. This model can be simulated in MATLAB and the parameters
can be used for the block diagram in Simulink. We are providing a x(n) as an input for memory
effect and y(n) is the output here.
1. The DPD derivation begin as follows:
Assume a memory polynomial form is f (X1, X2…Xn) for the non-linear Power Amplifier operator
𝑦𝑀𝑃(𝑁) = ∑ ∑ 𝑎𝑘𝑚
𝑀−1
𝑀=0
𝐾−1
𝐾=0
𝑥(𝑛 − 𝑚)|𝑥(𝑛 −𝑚)|𝑘(7.1)
Where,
x is the PA input
y is the PA output
akm are the PA polynomial coefficients
M is the PA memory depth
K is the degree of PA non-linearity
n is the time index
2. Inverse non-linear function for DPD, f-1(X1, X2…Xn)
𝑥𝑀𝑃(𝑁) = ∑ ∑ 𝑑𝑘𝑚
𝑀−1
𝑀=0
𝐾−1
𝐾=0
𝑦𝑠𝑠(𝑛 − 𝑚)|𝑦𝑠𝑠(𝑛 − 𝑚)|𝑘
Here, output y(n) is being normalized by G,
53
ys(n) = y(n)/G
yss(n) = ys (n+offset) [offset is a fixed +ve integer]
However, the PA has significant memory before obtaining the coefficients it is essential to offset
y. The predistrotion is one of the reason for the positive delay of the amplifier. Depending on the
time shift of the amplifier output, we can obtain the coefficient, but no model can control a negative
delay. Additionally, an amplifier output works as an input to obtain the DPD coefficient. However,
we created an equation yss(n) = ys(n+offset), upon which to base the coefficient derivation, where
output x act to the input y offset [6].
3. In figure 7.6, we rewrite equation 7.1 as a set of system of linear equations. First, solve
DPD coefficient, dkm then here are the linear equations for the over-determined system [6]:
Figure 7.6 Equation 7.1 Rearranged As a System of Linear Equations in Matrix Form [6]
The variable p= the number of computation samples. However, this is an over-determined system
where the product value of K*M is much less than the value of p. The values of x and y are knowns.
The value for K and M have chosen based on the 13.56 MHz E- Class PA. Where, we chose M
=1, K = 5 and K*M = 1*5= 5, complex coefficients.
Where, dkm = Yss/X
54
Figure 7.7 The Real and Imaginary Components of The Derived DPD’s Complex Coefficients.
7.4 Evaluating the Static-Coefficient DPD Design
We have created a static model for DPD along with our 13.56 MHz power amplifier subsystem to
evaluate the design where we firstly implemented equation 7.1 which is straight forward to
represent equation 7.1 in MATLAB as shown in Table 1.2 [6],
55
Table 1.2 Equation 7.1 Can Be Implemented Using Either MATLAB Code or Blocks in
Simulink
Equation
7.1 in
symbolic
form
𝑥𝑀𝑃(𝑁) = ∑ ∑ 𝑑𝑘𝑚
𝑀−1
𝑀=0
𝐾−1
𝐾=0
𝑦𝑠𝑠(𝑛 − 𝑚)|𝑦𝑠𝑠(𝑛 − 𝑚)|𝑘
Equation
7.1 in
MATLAB
code
Equation
7.1
In
Simulink -
block
diagram
form
56
Figure 7.8 DPD Verification Model
Figure 7.9 PA Behavior With DPD (1)
As our PA works in 13.56MHz, PA is working in 26.47 dbw. This plot shows the importance of
modeling non-linearities. All DPD variations adequately correct for in band distortion but a purely
linear DPD is not sufficient for reducing spectral regrowth.
57
Figure 7.10 PA Behavior With DPD (2)
After predistortion technique we obtained 16.07 dbw.
7.5 Extend static DPD design to adaptive
However, our static DPD model shows promising results, we implemented an adaptive test bench
where the matrix inverse math allows to solve an over-determined system of equations is not
possible to implement in hardware.
To implement an adaptive DPD, we used ILA as shown in figure 7.10. This design is summation
of DPD subsystem and Coefficient Calculation subsystem. There is no inverse matrix method is
the proposed model [6].
58
Figure 7.11 Our Adaptive DPD Test Bench
To obtain DPD coefficients we need coefficient subsystem represents power amplifier input and
output. By this DPD coefficient, the predistored output can be obtained. These DPD coefficient is
uniform and allows the coefficient to a memory polynomial [6].
Figure 7.12 shows the coefficient computation subsystem as follows
Figure 7.12 Coefficient Computation Subsystem Model [6]
Adaptive DPD model has two types of algorithm-
• For learning the coefficients and
• Other one for implementing them.
In figure 7.12 shows the summation of non-linear prod and the coefficient computation
subsystems.
59
Figure 7.13 The Nonlinear Prod Subsystem Implements The Non-Linear Multiplies [6]
The DPD coefficients, dkm, are computed on-the-fly using one of two possible algorithms, the LMS
algorithm or Recursive Predictor Error Method (RPEM) algorithm as shown in Figures 7.14 and
7.15 respectively [6].
Figure 7.14 The LMS-Based Coefficient Computation is Relatively Simple In Terms of The
Required Resources [6]
60
Figure 7.15 The RPEM-Based Coefficient Computation is Far More Complex Than The LMS
[6]
To form an error signal, both LMS and RPEM based coefficient computation algorithm is required.
The difference of the error signal is all about the measured power amplifier input value and
estimated power amplifier input value. Both algorithm tends to push error to zero and makes the
possible convergence on the DPD coefficient. The output behavior of adaptive DPD with PA
output and only PA output have been shown in Figure 7.16. The RPEM method shows remarkably
better results in both spectral growth reduction and rate of convergence.
61
Figure 7.16 Adaptive DPD Design Output Behavior
For Adaptive DPD Design PA gain is 110 dbW at the expense of significant spectral regrowth
and in-band distortion.
62
CONCLUSION
High efficiency and linear PAs are essential components in wireless communications system.
There is however a tradeoff between efficiency and linearity: high efficiency PAs behave
nonlinearly and linear PA present low efficiency. Due to fulfill the efficiency and linearity
requirements, DPD is often used. This thesis has contributed to different aspects of the linearization
of PAs using DPD.
The synthesis of a predistorter consists mainly in two tasks: the selection or design of models to
be used in the predistorter and the identification of the parameters of the model. For the parameter
identification, techniques such as pth-order inverse, ILA and DLA have been proposed in the
literature. Among them, ILA is the most widely used because it reduces the identification process
to an iterative inverse-modeling process. But the adoption of ILA introduces a critical issue known
as gain normalization. In the literature, there is not a clear consensus on how the normalization
gain must be computed. We have investigated this issue and have shown that normalization gain
affects the linearized PA’s output power. Consequently, it computes an extra degree of freedom
which should be selected carefully. To overcome this issue, we have proposed a variation of ILA
that terminate the normalization gain block and uses the desired output response as input to the
predistorter. Experiments showed that the proposed ILA does a better control of the output signal
from the linearized PA, since this signal depends on the input signal to the predistorter, not on the
normalization gain’s accurate selection.
An important issue encountered in DPD has been that the output signal from the predistorter that
linearizes the PA is unknown. For this reason, the parameters of the predistorter are identified
using techniques such as ILA and DLA. To overcome this issue, we have proposed a novel
parameter identification technique based on ILC. To this end, we have designed an ILC scheme
for PA linearization. Experimental results showed that the proposed ILC scheme can identify the
best linearizing input signal. The experiments also showed that proposed calculating technique can
obtain better linearization performance than ILA and DLA when high levels of measurement noise
are present and when the PA is in high condensation. The ILC scheme and the identification
technique presented here have the capable to enable the design of model structures for digital
predistorters.
63
The introduction of dual-input transmitter architectures has introduced new challenges to digital
predistortion. In this work, we have studied linearization aspects of dual-input Doherty PAs and
identified some of the associated challenges. We have experimentally shown that the large
bandwidth of the control signals and the limited sampling rates used in the linearization are two
the main factors that limit their linearity performance. We have addressed this issue with a novel
linearization scheme that gives a better linearity compared to the existing schemes for a given
sampling rate.
In summary, we have developed new concepts and methods for DPD linearization of PAs. These
contributions are expected to make a broad impact on the realization of energy efficient and linear
radio transmitters for emerging and future wireless systems.
64
FUTURE WORK
We implemented the software part in this thesis. However, the noise output is a bit higher, in future
it is implementable the hardware part. We have a PXIe 1075 chassis in our RF/Microwave lab.
Firstly, we must measure the strength and give a physical form of our E-class PA and then PXIe
1075 chassis and spectrum analyzer could help us to compute the AM/AM and AM/PM. Besides
we could get the linearize behavior of our PA after Digital Pre-Distortion.
Moreover, Linearized E-class PA’s are very suitable for 5G communications. They can be
designed in an extreme dense size to minimize the cost. Moreover, mm-wave power amplifiers
have high efficiency at peak and the amount of power withdraw that are relative to those while
operates at low frequencies. The greater number of RF signal bandwidth in 5G Frontends mean
that DPD (Digital Predistortion) will be completely different from their 4G predecessors.
65
REFERENCES
[1] Acar, M., Annema, A. and Nauta, B. (2007). Analytical Design Equations for Class-E Power
Amplifiers. IEEE Transactions on Circuits and Systems I: Regular Papers, 54(12),
pp.2706-2717. Retrieved from
https://www.researchgate.net/publication/3451728_Analytical_Design_Equations_for_Cl
ass-E_Power_Amplifiers
[2] Aslan, M., Atalar, A., Reyhan, T., Ilday, F., (2009). Adaptive Digital Predistortion for Power
Amplifier Linearization. Master’s Thesis, Bilkent University, Turkey. Retrieved from
http://repository.bilkent.edu.tr/handle/11693/14828
[3] Chani-Cahuana, J. (2015). Digital Predistortion For The Linearization of Power Amplifiers.
Department of Signals and Systems, Chalmers University of Technology, Sweden.
Technical Report R008/2015 ISSN 1403-266X. Retrieved from
https://www.chalmers.se/en/centres/ghz/publications/Documents/Lic_thesis_18-
%20Digital_predistrotion_for_the_linearization_of_power_amplifiers.pdf
[4] Erdogd, G. (2012, June). Linearization of RF Power Amplifier By Using Memory
Polynomial Digital Predistortion Technique. Master’s Thesis, The Graduate School of
Natural And Applied Science of Middle East Technical University, Ankara, Turkey.
Retrieved from
http://citeseerx.ist.psu.edu/viewdoc/download?doi=10.1.1.633.1178&rep=rep1&type=pdf
[5] Morgan, D., Ma, Z., Kim, J., Zierdt, M., & Pastalan, J. (2006). A Generalized Memory
Polynomial Model for Digital Predistortion of RF Power Amplifiers. IEEE Transactions
on Signal Processing, 54(10), 3852-3860. doi:10.1109/tsp.2006.879264.
[6] Schutz, K. Adaptive DPD Design: A Top-Down Workflow. MathWorks. Retrieved from
https://www.mathworks.com/company/newsletters/articles/adaptive-dpd-design-a-top-
down-workflow.html
[7] Raviv, R. (2004, March). Nonlinear System Identification and Analysis with Applications to
Power Amplifier Modeling and Power Amplifier Predistortion. PhD Thesis, School of
Electrical and Computer Engineering, Georgia Institute of Technology, USA. Retrieved
from
https://pdfs.semanticscholar.org/1700/1e4e2530695c166d0f1809d6893ed3eb2ae1.pdf?_g
a=2.88895826.853710321.1544030109-181856262.1541694594
[8] Sezgin, I. C. (2016). Different Digital Predistortion Techniques for Power Amplifer
Linearization. Master’s Thesis, Lund University, Sweden. Retrieved from
https://www.eit.lth.se/sprapport.php?uid=954
[9] Singerl, P. (2006, December). Complex Baseband Modeling and Digital Predistortion for
Wideband RF Power Amplifiers. European Association for Signal Processing. PhD
66
Thesis, Graz University of Technology, Austria. Retrieved from
https://theses.eurasip.org/media/theses/documents/singerl-peter-complex-baseband-
modeling-and-digital-predistortion-for-wideband-rf-power-amplifiers.pdf
[10] Teikari, I., (2008). Digital Predistortion Linearization Methods For RF Power Amplifiers.
Doctoral Dissertation, Department of Micro and Nanosciences, Helsinki University of
Technology, Finland. ISBN 978-951-22-9545-6, ISBN 978-951-22-9546-3 (PDF), ISSN
1795-2239, ISSN 1795-4584 (PDF). Retrieved from
http://lib.tkk.fi/Diss/2008/isbn9789512295463/isbn9789512295463.pdf
[11] Yusop, Y., Saat, M. S., Husin, S. H., Nguang, S. K., & Hindustan, I. (2016). Design and
Analysis of 1MHz Class-E Power Amplifier for Load and Duty Cycle Variations.
International Journal of Power Electronics and Drive Systems (IJPEDS), 7(2), 358.
doi:10.11591/ijpeds.v7.i2.pp358-368.